Mapping soil particle-size fractions based on compositional balances

Mapping of soil particle-size fractions (PSFs) combined with log-ratio transformation has been widely employed, particularly for isometric log-ratio (ILR). Regression kriging (RK), as a hybrid interpolator, represents a method to enhance prediction accuracy. However, different choices of ILR balance yield distinct transformed data. It remains unclear and lacks a comparison as to whether these results exhibit robustness when employing RK modeling. In this study, we compared the performance of four modelling approaches-generalized linear model (GLM), random forest (RF), and their hybrid models (i.e., GLMRK and RFRK). These models were applied to three ILR transformed datasets based on different balances, resulting in a total of 12 models, in the upper reaches of the Heihe River Basin, China. The results indicated that RF tended to provide more accurate predictions of soil PSFs, while GLM was better at producing predictions with less bias. The study recommends the use of RK, as it was found to broaden the value ranges of predictions, adjust bias, and enhance accuracy, especially when applied in conjunction with RF models. Furthermore, prediction maps generated from RK unveiled finer details of soil sampling points. The choices of ILR balance resulted in varying data distributions for the components of sand, silt, and clay. These components tended to cluster at approximately 120 degrees in three groups, thereby indicating that even components with small content also exert significant influence on soil compositions. Rather than solely focusing on the relative abundance of components, this study suggests that the alignment of ILR components with a normal distribution is crucial for better model performance, especially for the first ILR component. Opting for the most abundant component as the first permutation may not always lead to optimal results for soil PSF mapping. This study provides insights into the role of data distribution and ILR balance selection in soil PSF mapping with transformed data.